This week at CES 2026 in Las Vegas, Tensor announced the open-source release of OpenTau (τ)—a training toolchain built to speed up and simplify the development of Vision-Language-Action (VLA) foundation models, an emerging core component of “Physical AI” systems.

If you work in autonomous driving, robotics, or embodied AI research, you already know the direction the field is moving: models that don’t just see or talk, but can perceive the world, reason about it, and take actions—all within a single multimodal foundation model. That is the promise of VLA.

OpenTau is Tensor’s attempt to make training these models more reproducible, accessible, and scalable—and to push the broader ecosystem forward by putting advanced training infrastructure into the open.

Why VLA Matters for Physical AI

Physical AI is fundamentally different from purely digital AI. The real world is messy: environments change, sensors are noisy, and actions have consequences. Building systems that can operate reliably in this setting requires models that can combine:

• Vision (what’s happening in the environment),
• Language (instructions, goals, context),
• Action (what the system does next).

VLA models aim to integrate all three into one foundation model so intelligent systems can interpret inputs, plan, and execute actions—supporting use cases like robotic manipulation, navigation, and autonomy in driving.

What OpenTau (τ) Is

At Tensor, we believe meaningful progress in Physical AI requires transparency. OpenTau is Tensor’s open-source training platform for frontier VLA models, designed to make large-scale training workflows more practical outside of closed or proprietary environments.

Tensor’s stated goal with open-sourcing OpenTau is to enable more scientific transparency and independent validation—making it easier for researchers and developers to reproduce results, experiment with new training strategies, and build on top of a common toolchain.

What’s Inside: Key Training Capabilities

OpenTau brings several state-of-the-art capabilities into an open-source toolchain, including:

• Co-training on an adjustable mixture of heterogeneous datasets
  Train across multiple datasets with tunable proportions—useful when combining different domains, sensor setups, or task formats.
• Discrete action modeling for faster Vision-Language Model (VLM) convergence
  A strategy intended to improve training speed and stability as the model learns to connect perception and language to action outputs.
• Knowledge insulation between the VLM backbone and the action expert
  Architectural or training approaches that separate responsibilities between components—helpful when scaling or refining action behavior without destabilizing the underlying backbone.
• VLM dropout techniques to reduce overfitting
  Regularization designed to improve generalization, especially when datasets are limited or skewed.
• A reinforcement learning pipeline purpose-built for VLA models
  A training path for aligning action behavior through RL methods tailored to the VLA setting.
• And more

Overall, the emphasis is on reproducibility and extensibility—enabling teams to test advanced training strategies that can otherwise be difficult to implement without significant internal infrastructure.

How to Get Involved

Tensor is inviting the community—researchers, developers, and builders—to explore OpenTau, contribute improvements, and extend it for new VLA model work.

You can find the repository here:

OpenTau on GitHub: https://github.com/TensorAuto/OpenTau

Ways to participate include starring the repo, forking the codebase, opening issues, and contributing PRs—especially around dataset integrations, training recipes, evaluation workflows, and new VLA experimentation.

‍